From question to sketch
For a long time the problem stayed a question: a system can retrieve, but it does not remember. In mid-2024 it became a sketch β the first concrete attempt to describe, on paper, what a memory subsystem for an AI system would have to contain. We gave the sketch a working name, ZenBrain, mostly so we could argue with it.
The starting commitment was deliberate and slightly unfashionable: rather than designing from what is convenient to build, design from what is known about how memory is organised, and only then ask what is implementable. The literature here is old and unusually solid.
The findings the sketch had to honour
Four results did most of the shaping.
Tulving (1972) separates episodic memory β concrete experiences tied to a time and place β from semantic memory, the abstract knowledge distilled out of them. "We discussed the budget on Tuesday" is stored differently from "budgets need owners." A serious memory needs both, and a way to move from the first to the second.
Ebbinghaus (1885) described the forgetting curve: memories decay exponentially, but each well-timed review flattens the curve. Forgetting, on this view, is not a defect to be engineered away. It is the mechanism that keeps the important separable from the incidental.
Hebb (1949): repeated co-activation strengthens a connection. Association is something that accrues with use, not a static link drawn once at write time.
Sleep consolidation (in the tradition of Stickgold and Walker) suggests that some of the most important memory work happens offline β replaying the day, strengthening what matters, weakening the rest. That implies a memory system needs a background process, not just a read and a write path.
What the first sketch proposed
Out of those commitments came an early shape: not one memory store but several, each with a distinct role β an active working focus, a short-term session context, an episodic record of concrete events, a long-term store of distilled knowledge, a procedural memory for learned routines, and a small set of pinned, always-present facts. Consolidation would move information between them over time rather than at the moment of writing.
In mid-2024 this was a diagram and a stack of references, nothing more. There was no code, no benchmark, no claim that it would work. What it did have was a property we cared about: every part of it could be traced to a documented mechanism and, in principle, measured. Whether the pieces would cohere into something better than a vector store was exactly the open question β and the reason to start building.
Next in this series: Seven Layers: A Memory Model Takes Shape.